Relation extraction for Assamese, a morphologically rich Indo-Aryan language spoken by roughly 15 million people, has been held back by the absence of an annotated corpus, a trained model, and any standardised evaluation protocol. We address all three gaps in this work. We first introduce AREX, the first manually annotated Assamese relation extraction dataset, comprising 1,927 triples across ten semantic relation categories drawn from biographical Wikipedia text, with inter-annotator agreement of κ=0.84. We benchmark four fine-tuned multilingual transformer encoders together with one instruction tuned large language model under a controlled five seed protocol, and we find that task specific fine tuning of the smaller multilingual encoders outperforms zero shot prompting of the substantially larger LLM, with mBERT reaching a macro F1 of 0.772 ± 0.012 against Llama-3.1-8B at 0.719 ± 0.016. To exploit the complementary strengths of the four fine tuned models, we then propose a Relation Guided Adaptive Ensemble that computes per relation weights from validation F1-scores and aggregates the corresponding probability distributions at inference. The ensemble achieves a macro F1 of 0.782 ± 0.007, significantly improving on mBERT under a paired t-test at p=0.049 and reducing across seed variance by 40%. An ablation study confirms that both the F1-guided weighting and the per relation granularity are individually necessary, with uniform weight averaging performing worse than the strongest single model. The extracted triples are serialised into an RDF knowledge graph of 2,184 edges and 1,467 entities, queryable through SPARQL and enriched with temporal attributes for time anchored relations. A manually verified stratified sample of 50 triples yields a triple level precision of 84.0% with a 95% Wilson confidence interval from 71.7 to 91.2 percent and verification agreement κ=0.86. AREX, the trained models, and the constructed knowledge graph collectively provide the first end-to-end reproducible baseline for Assamese relation extraction and demonstrate a practical pipeline for extending information extraction to other morphologically rich low resource languages.
Melanoma is one of the most life-threatening forms of skin cancer, where early detection improves survival outcomes. Despite recent advances in Deep Learning (DL) for dermoscopic image analysis, existing approaches face critical limitations, including reliance on image-only inputs, inability to incorporate clinical metadata, and unstable performance on imbalanced datasets observed in real-world screening. Furthermore, many architectures are computationally heavy, limiting their suitability for clinical deployment. To address these gaps, in this paper, we propose a lightweight multi-modal diagnostic framework (BNDrop-Xnet) that integrates dermoscopic image features with clinical metadata (age, sex, and lesion location) through a weighted fusion strategy. Our framework consists of four main stages: (1) data preprocessing, including normalization, stratified sampling, and SMOTE-based class balancing; (2) image model creation using a four-block Convolutional Neural Network (CNN) architecture with Conv2D, Batch Normalization, MaxPooling, and Dropout layers; (3) metadata model construction with impact-factor computation and Multi-Layer Perceptron (MLP)-based learning of diagnostic feature contributions; and (4) ensemble integration through a weighted fusion and decision threshold optimization. It was evaluated on the ISIC 2020 dataset and achieved 90% accuracy, 80% precision, and a 73% F1-Score, outperforming several state-of-the-art models. These results demonstrate the effectiveness of combining visual and clinical cues for robust melanoma screening. BNDrop-XNet offers an efficient, interpretable, and meaningful solution for real-world deployment in resource-limited environments.
Zero-shot and few-shot learning methods offer effective solutions to data scarcity in Natural Language Processing (NLP), especially for under-resourced languages like Arabic and for mixed-language (code-switched) text. However, their capabilities in such challenging settings remain underexplored. In this work, we present the first systematic comparison of two prompt-based few-shot learning approaches; Pattern-Exploiting Training (PET) and LM-BFF (Better Few-Shot Fine-Tuning) on Arabic and Arabic-English code-switched classification tasks (Named Entity Recognition and Sentiment Analysis). Our experiments show that LM-BFF achieves a new state-of-the-art result on an Arabic-English sentiment dataset (ArEnSA), outperforming the previous best F1-score by approximately 1.5% points, and consistently surpasses PET in sentiment classification, particularly under code-switched conditions. Meanwhile, PET remains a strong baseline for structured tasks such as Arabic NER. These findings highlight the efficacy of prompt-based fine-tuning in low-resource, multilingual NLP, and underscore the potential of such techniques for underrepresented languages and code-switched text.
Constructing high-quality grounded question-answer (QA) datasets remains a major bottleneck in developing reliable retrieval-augmented generation and question-answering systems, particularly for domain-specific and low-resource corpora. Although large language models (LLMs) can automatically generate QA pairs, the quality of generated datasets depends heavily on prompt design and reliable validation mechanisms. This paper presents a controlled comparison of Single-Path Thematic Prompting (SPTP) and Multi-Path Thematic Prompting (MPTP) for automated grounded QA dataset construction from Indonesian translations of the Quran. Within each generator condition, thematic path structure is treated as the manipulated prompt variable while the task instructions, output schema, grounding constraints, and validation criteria remain fixed. The primary full-corpus experiment generates QA pairs from 4685 thematically indexed verses using GPT-5.2 and validates them using DeepSeek-V3.2, Mistral-Large-3, and Llama-4-Maverick. To directly evaluate cross-generator portability, we additionally compare GPT-5.2 and GPT-OSS-120B on a four-way matched subset of 2571 verses. Outputs from both generators under SPTP and MPTP are evaluated using the same open-weighted validator ensemble consisting of Gemma4:31B, Granite4.1:30B, and Qwen3.6:35B. Validation adopts two acceptance criteria: majority consensus (at least two of three validators support the generated answer) and strict consensus (all three validators support the generated answer). On the full proprietary track, MPTP significantly improves strict consensus over SPTP (p=.00415, McNemar’s test), while majority acceptance remains similarly high. On the matched cross-generator subset, GPT-5.2 achieves majority acceptance rates of 97.3% under SPTP and 97.5% under MPTP, while strict acceptance increases from 86.4% to 88.7% (p=.00274). GPT-OSS-120B achieves majority acceptance rates of 94.3% and 94.7%, respectively, whereas strict acceptance remains identical at 80.4% under both prompting conditions (p=.966). Majority-level acceptance decisions agree between the two generators for 93.4% of SPTP items and 93.5% of MPTP items. The resulting dataset contains 4644 automatically validated grounded QA pairs covering 99.1% of the attempted verses. The results provide evidence that the overall generation–validation framework is applicable across different generator families. However, the magnitude of the benefit produced by multi-path thematic prompting is generator-dependent rather than universally invariant.
Emojis are present in user-generated content, mainly on the Internet. This fact makes emoji analysis an important aspect of information retrieval from natural language texts. This work explores the potential of emojis as indicators for rating prediction of online product reviews. As a case study, we analyze the reviews of apps on the Google Play Store. The results indicate a significant presence of emojis in the comments, particularly in extreme ratings (1 and 5 stars), consistent across languages. We conducted a quantitative evaluation of emoji-only, text-only, and text+emoji configurations using eXtreme Gradient Boosting (XGBoost) and multilingual BERT models. In 102,118 reviews from 13 European locales, the Multilingual ModernBERT (mmBERT) text+emoji model achieves MAE 0.8131 and F1-micro 0.4321, while the XGBoost text+emoji model achieves MAE 0.9749 and F1-micro 0.3702 with a shorter training time. Emoji-only models remain competitive, with XGBoost achieving MAE 0.9773 and F1-micro 0.3629, which becomes a lightweight solution for rating prediction, especially at extreme ratings. We also analyze language and country conditioning, noting stabilization and cross-lingual variability. Training efficiency per fold is reported for deployment considerations.
Current document KIE systems are typically optimized for end-task accuracy, but deployment also requires evidence for predicted links, complete processing of long documents, and stable joint SER+RE training. In this work, we study KIE reliability through three requirements: inspectable decision evidence, coverage completeness under long-document inputs, and optimization stability in joint SER+RE learning. We instantiate this perspective with a unified framework consisting of three modules. Global Entity Contextualization (GEC) introduces global entity interaction before pair scoring and exposes entity-level attention patterns as supporting evidence for inspection. Coverage-Preserving Inference (CPI) extends sliding-window chunked decoding to evaluation with deterministic SER merging and cross-chunk relation reasoning, reducing truncation-induced silent failures. Span-Conditioned Relation Initialization (SCRI) initializes relation learning on realistic predicted spans, reducing early-stage gradient conflict in joint optimization. On HUST-CELL (ICDAR 2023 SVRD), the framework improves CompScore from 54.88% to 59.93%, surpassing the reported competition champion (56.45%) by 3.48 points. Beyond benchmark gains, the method provides inspectable relation evidence and robust long-document behavior.
Traditional feature extraction methods have struggled to capture nuanced emotional expressions while, transformer-based approaches although more effective often depends on computationally expensive fine-tuning. To address these issues, this research explores the use of Sentence Transformer embeddings as fixed feature extractors for emotion classification tasks. Experiments were conducted on ISEAR benchmark dataset, where ML classifiers were trained on embedding generated from four base pre-trained sentence transformers. The evaluation was performed using both holdout and Stratified 5-Fold cross-validation strategies. Among the tested approaches, using stratified 5-fold cross validation, the E5-base v2 (BERT Based) embedding model combined with a support vector classifier achieved an accuracy of 69.90%, outperforming proprietary embedding baselines without the need for fine tuning. Statistical validation performed using paired t-tests further confirmed that performance differences between embedding models are significant and not attributed to random data partitioning. Additionally, empirical finding indicates that embedding fusion of BERT based pre trained sentence transformers (BGE, E5, and GTE) performance is marginally comparable to that of classifier trained on individual embedding using same evaluation strategy. These results demonstrate that open-source sentence embeddings constitute a computationally efficient and competitive alternative to fine-tuned transformer models for emotion classification tasks.
Text representation is fundamental to Natural Language Processing (NLP), as it transforms human language into numbers (vectors) that machines can interpret. The quality of the representation directly determines the effectiveness of downstream tasks such as sentiment analysis, translation, and question answering. This survey adopts a multidimensional evaluation framework to analyze text representation techniques, focusing on their evolution, performance, and applications across various NLP tasks. It proposes a text representation taxonomy that categorizes embedding methods based on their granularity for specific use cases. Various evaluation metrics often used in assessing embedding performance were discussed. In addition, the survey reviewed significant data catalogues relative to their applicable NLP task, which is crucial in generating an embedding. It provides a critical analysis of emerging trends, while offering insights into their potential to address scalability and domain adaptation challenges. The paper’s unique contribution lies in its comprehensive projection of historical developments, current advancements, and future directions, providing researchers and practitioners with a simple yet valuable understanding of the field.
This study presents a systematic literature review examining the evolving role of Natural Language Processing (NLP) in audio and music artificial intelligence between 2021 and 2025. While early research primarily treated language as an auxiliary component supporting signal-based modelling, recent developments indicate a fundamental conceptual shift in which language functions as a central mechanism for meaning construction, mediation, and explanation across modalities. Following the PRISMA framework, this review analyses 139 peer-reviewed journal articles to identify how linguistic functions are formalized as computational tasks in speech, audio, and music systems. The findings reveal a progressive transition from acoustic-centric approaches toward language-driven representations that account for semantic, pragmatic, and contextual dimensions of meaning. In speech processing, utterances are increasingly modelled as linguistic and cognitive structures rather than mere transcriptions. In music-related research, lyrics and linguistic descriptions emerge as primary carriers of musical meaning, enabling genre interpretation, affect modelling, and language-based music control. At the intersection of NLP, audio, and music, language operates as a semantic mediator aligning acoustic representations with musical structures. Beyond mapping this evolution, the review identifies persistent methodological and conceptual challenges related to non-literal meaning, cultural variability, evaluation inconsistency, dataset bias, and interpretability. Overall, this review argues for positioning language as a foundational axis for future Audio and Music AI research and contributes to a structured basis for more semantically grounded, context-aware, and explainable systems.
Arabic smishing detection is difficult because short SMS messages combine phishing intent, Arabic morphology, dialectal variation, code-mixed tokens, and deliberate obfuscation. This study presents a revised, application-oriented framework for Arabic smishing detection built around a source-audited corpus, Arabic-specific preprocessing, a CNN-BiGRU-attention classifier, and post hoc explanation. The corpus contains 5576 Arabic SMS-style messages drawn from public smishing reports, institutionally authorized legitimate messages, and synthetic augmentation used only for training. Synthetic instances account for 34% of the augmented training resource (1,896 of 5576 examples), and all resampling is performed after train/validation/test separation to reduce leakage. Preprocessing combines Unicode normalization, orthographic normalization, clitic-aware Farasa segmentation, and domain-aware handling of URLs, numerals, and mixed-script tokens. The hybrid model is evaluated on a single locked prediction file against classical text classifiers, recurrent and convolutional baselines, and Arabic transformer baselines based on AraBERT and CAMeLBERT, with leakage-aware splitting, training-fold-only augmentation, bootstrap confidence intervals, paired significance testing, and component ablation. On the locked held-out test set (n = 1,000; threshold τ = 0.50), the hybrid model achieved 98.60% accuracy, 99.39% precision, 97.80% recall, an F1 score of 98.59%, and ROC-AUC of 0.992; every reported metric, the confusion matrix, and the headline figures are now regenerated from the same prediction file. Paired McNemar tests with Holm correction show that the hybrid model is significantly stronger than weak classical baselines but statistically indistinguishable from the strongest neural and Arabic-transformer baselines, so the contribution is framed as an operationally credible, precision-oriented alternative rather than a state-of-the-art claim. Explainability is reframed around quantitative validation of LIME through fidelity, stability, human-rationale agreement, and analyst utility rather than anecdotal inspection. Error analysis links false positives and false negatives to Arabic morphology, dialectal paraphrase, Arabizi, and adversarial Unicode obfuscation. The contribution of the paper lies not in proposing a new neural architecture, but in specifying a more rigorous Arabic smishing pipeline that combines dataset governance, reproducible evaluation, and operationally interpretable analysis.
Background Language endpoints serve as meaningful clinical outcomes in testing for neurological conditions (e.g. dementia). Assessments can be conducted using paper/pencil tests or through digital outcomes using natural language processing (NLP). The accuracy of many NLP models across languages have been reported with varied methodologies and levels of transparency. This warrants replication using a standard methodology. This study aimed to develop a standard methodology across languages for evaluating POS taggers and to report initial results. Methods 105 linguists tagged language features across 58 language varieties using a short, standardized translated passage. When two linguists were available, we calculated inter–rater reliability using Cohen’s kappa. We obtained the tokenization and parts of speech tags across the 52 language varieties from available models from four accessible and reproducible NLP libraries. We calculated the percentage of matches between NLP models’ and linguists’ tokenization. We compared the NLP models’ parts of speech tags to the linguists’ using Matthew’s Correlation Coefficient. Findings Twenty–five language varieties had strong performance (high reliability between linguists and strong agreement between NLP models and linguists). Eleven yielded unclear results and 16 had relatively poor performance. Interpretation Next steps for each language variety are suggested according to our results. Gathering reliable ‘ground truth’ annotations from multiple linguists is needed for some languages. Model improvement is needed for language varieties that yielded poor performance. For language varieties with promising results, replication using longer, more ecologically valid samples is warranted. Accordingly, our methods and data are transparently reported, facilitating next steps across languages.
In today's healthcare environment, understanding patient emotion is an invaluable measure for evaluating service quality, finding significant loopholes, and improving therapeutic outcomes. Yet, recently devised transformer-based models, such as GPT-3.5, Health-LLM, and Polaris, while brimming with promise for advancing clinical NLP conceptually, are frequently unable to handle the kind of terminology, negations, and simple statements that are ubiquitous in a domain, let alone clear and instantaneous explanations. Such models are too inefficient for deployment in actual healthcare system settings, lack sophisticated emotional reasoning, and are often poor performers when dealing with ambiguous input. Enter SENTICARE-GPT, a novel model for patient sentiment analysis that emerges from the combination of clinical natural language processing, emotion detection, and generative transformers. It starts from a Clinical Feedback Preprocessor (CFP+), which expands abbreviations and remembers context-sensitive phrases while normalizing medical language. Further, boasts of a Cross-Fused GPT Decoder (CF-GPT) with Dual Cross Attention that consolidates the semantic and affective signals every transformer layer. The Contrastive Clinical Sentiment Adapter (CCSA) mentioned here aids in categorization under ambiguous feedback by a mere recognition of short sentiment changes. As a multi-agent classifier, the Sentiment Intensity & Risk Estimator (SIRE) forecasts urgency level, emotional intensity, and polarity. Finally, the Explainable Sentiment Tracer (X-ST) permits healthcare professionals to make transparent sentiment decisions by providing phrase-level justifications based on integrated gradients and generating justification modules. The experimental evaluations show that SENTICARE-GPT is very powerful for real-time interpretable healthcare sentiment analysis, ranking higher than existing state-of-the-art models in sentiment accuracy (91.6%), urgency recall (90.7%), and explainability satisfaction (4.6/5).
Triplet extraction-aspect, opinion, and sentiment detection for each product review is a valuable contribution to fine-grained sentiment analysis. Though this task is extensively studied for English and other large languages, triplet extraction for Bengali was never addressed by any prior work. In this paper, we introduce the neural system for Bengali aspect-opinion-sentiment triplet extraction leveraging the harmony between BanglaBERT embeddings and BiLSTM structures in a multi-stage pipeline. To enable this task, we present the BPR Corpus, a large manually annotated Bengali dataset with fine-grained triplet labels, filling a critical resource gap in low-resource language research. Our approach meticulously extracts aspect terms, extracts corresponding opinion words, and determines sentiment polarity for every pair, achieving F1 scores of more than 0.82 on all subtasks. Exhaustive experiments on genuine product reviews corroborate the generality and scalability of our model. This study establishes a new benchmark for sentiment analysis for Bengali and provides data and methodological foundations for under-resourced languages for future research.
Code comment generation aims to produce concise natural-language comments that help developers understand and maintain source code. Recent work has expanded the input to include information outside the target method, but most approaches still treat context as a single pool of evidence. This is limiting in multi-intent summarization, because a comment that explains what a method does may require different evidence from a comment that explains how to use the method or why it exists. We propose IntentContextCS, a framework for multi-intent code comment generation that selects context according to the intended documentation goal. The framework combines three selection signals: token overlap between the target method and context types, semantic similarity based on CodeBERT embeddings, and reasoning-based context ranking using large language models. These signals are used to construct an intent-specific policy over five context types: callers, callees, sibling methods, class attributes, and project-level information. On the CodeSearchNet Java dataset, the empirical analysis shows that intent categories have different context profiles. For example, caller contexts are most useful for usage-oriented comments, whereas callee and sibling contexts are more informative for comments about functionality and implementation. In the generation experiments, IntentContextCS improves METEOR by 8.9%, USE by 3.5%, and SIDE by 3.0% over the strongest full-context LLM baseline, while using 54% fewer context tokens. These results suggest that, for code comment generation, selecting the right context is more effective than simply providing more context.
Current pipelines are disconnected, with generation systems generating fluent text but without strong evaluation and evaluation approaches that rely on heuristics or shallow proofreading, missing a coupling of rhetorical fluency and logical rigor. These limitations have been identified in terms of argumentative validity, domain adaptability, and alignment with the journal’s interdisciplinary impact/innovation focus. To solve those issues, a closed-loop paradigm was introduced where large language models are coupled with cognitive-logic veracity modules that operate in an adversarial training manner. In particular, the proposed approach switches between a generator, which is a transformer-based model fine-tuned on academic style, and an evaluator, which has been trained to distinguish flaws in the structure of the discourse, logical coherence, and alignment within the domain. The cognitive-logic step also operates to confirm rigorous logical inference, formal structure, and argument validity at each stage of iteration of the manuscript. Along with reinforcers as feedback, the generator optimizes to produce text that satisfies linguistic fluency and cognitive rigor. Evaluating the presented method on a corpus of text representative of student papers in a range of computing disciplines, significant 15%, 10%, and 8% positive differences were found in coherence, style, and timeliness, respectively, from generative-only baselines.
The article is potentially destined to examine the efficacy of Large Language Models (LLMs) in Swahili (standard dialect spoken in Tanzania), a relatively less privileged and Low-resource Language (LRL) that, to some extent, remains underrepresented in AI communication technologies. Despite the rapid growth in LLM use, Swahili users in Tanzania often encounter inaccuracies and unclear outputs, highlighting persistent challenges in model performance. The inaccuracies of LLMs in Swahili undoubtedly demonstrate the challenges associated with their use and effectiveness in such a language. In this study, participants from Tanzanian Swahili-speaking communities evaluated the models' outputs through usability tests. Findings reveal that apart from architectural limitations, data scarcity drives the ineffectiveness of the models. Frequent distortions, mostly in Swahili than in French confirms the need for broader multilingual inclusion in LLM training. The study highlights the imperative for inclusive AI development that empowers low-resource languages.
